Can large language models reliably audit accessibility in component libraries?

AI · 4 min read

Multiple vendor tools now wrap large language models to scan component stories and generate accessibility reports. These systems can suggest missing ARIA attributes, flag ambiguous alt text, and propose label copy based on component usage. Hooked into Storybook or a design system's documentation, they produce plain-language recommendations faster than a manual audit.

However, LLMs have limits. They sometimes miss context-specific semantic meaning — for example, whether an icon-only button truly requires a visible label or will consistently be supported by additional text in real product pages. They can also overconfidently recommend ARIA patterns that conflict with native semantics, creating brittle implementations.

Teams are finding the best approach is hybrid: use AI to triage and annotate potential issues, then let accessibility engineers and designers validate fixes. Embedding explainability — why a suggestion was made and citing patterns — reduces the risk of blindly applying AI recommendations and improves developer trust in the toolchain.